Can Large Language Models Do Research? What Theoretical Computer Science Benchmarks Reveal About AI’s Research Limits
Large language models continue to achieve impressive scores on mathematics, coding, and reasoning benchmarks. But solving well-defined problems is not the same as conducting research. A recent evaluation framework for theoretical computer science separates knowledge retrieval, paper understanding, construction, computation, and exploratory proving to measure how models perform on research-like tasks. Its central finding is cautious: no evaluated model independently solved an open problem, while machine-verifiable tasks produced the most reliable results. The study offers a practical framework for understanding where AI can assist researchers—and where human judgment remains essential.
1. High Benchmark Scores, but How Close Are We to “Doing Research”?
Large language model leaderboards are refreshed almost every few months. Scores in mathematical reasoning, code generation, and logical question answering continue to rise. Yet a closer look at many of these benchmarks reveals a common limitation: they primarily measure problem-solving ability, not research ability.
The difference can be illustrated with a simple analogy. Solving an exercise is like cooking from a recipe: the steps are already defined, and the result can be judged against an established standard. Research is different. There may be no predefined path, the answer may be unknown even to the person posing the question, and success cannot be guaranteed in advance.
Recently, a multi-institutional team from universities and research organizations published an evaluation covering the full field of theoretical computer science. Its central question was straightforward: if AI is asked to perform genuinely research-oriented work, what can it actually accomplish?
The initial conclusion was cautious. None of the most capable evaluated models could independently solve an unresolved open problem using their own abilities.
2. Why Use Theoretical Computer Science as the Testbed?
Evaluating AI’s research capabilities presents a fundamental trade-off. The closer a task is to real research, the harder it is to grade reliably. The easier it is to grade, the more likely it is to resemble a routine exercise.
Theoretical computer science offers a useful way to address this tension. It is not primarily concerned with repairing computers or writing application code. Instead, it studies the nature of computation itself.
One analogy is that programmers are like drivers, focusing on the engine and the interior of a vehicle, while theoretical computer scientists study transportation: how quickly can someone travel from one place to another, and are there destinations that cannot be reached by any route?
Several characteristics make the field particularly suitable for evaluating AI:
- A mathematical core: The field sits at the intersection of mathematics and computer science. Its questions are centered on computation, and its primary method is constructing proofs.
- Machine-verifiable results: The correctness of a proof can be checked by formal verification systems. This is similar to a Sudoku puzzle: finding a solution may be difficult, but checking whether a completed grid is valid is comparatively straightforward.
- A clear range of difficulty: The field includes everything from textbook-level theorems to open conjectures that have remained unresolved for decades. This makes it possible to map model capabilities with greater precision.
As mathematician Terence Tao has noted in interviews with academic publications, AI has a serious weakness across almost every field: it can produce errors that are difficult to verify. Mathematical proofs are different because their correctness can, in principle, be checked automatically. This is one reason why the most defensible achievements reported by AI companies often appear in mathematics.
3. Two Major Problems with Existing Benchmarks
1. Data Contamination and the “Memorization” Effect
Many traditional benchmarks draw questions from public datasets, competitions, or textbooks. These materials may already be present in a model’s training data. As a result, a model might be recalling a memorized answer rather than genuinely reasoning through the problem.
Distinguishing between these possibilities requires newly constructed questions that are not present in the model’s training corpus.
2. Grading Costs and Reliability
When a model produces hundreds of pages of purported proofs, reviewing every page manually is practically impossible. Some organizations have published hundreds of AI-generated mathematical manuscripts and categorized them by relevance, prompting public discussion about whether AI is overwhelming the field of mathematics. But determining how many of those manuscripts withstand rigorous verification requires a substantial amount of work in its own right.
Unless the grading problem is addressed, benchmark conclusions remain difficult to trust.
The evaluation discussed here adopted a clear strategy: use only tasks that machines can grade. The answer format must be automatically verifiable, avoiding the bottleneck of manual peer review.
4. A Research-Oriented Evaluation Framework
The evaluation created a 100-question test set organized around four types of research activity. These categories correspond to four common forms of work performed by theoretical computer scientists:
- Knowledge retrieval: Tests whether a model accurately understands classic results, definitions, and techniques, and whether its underlying knowledge is reliable.
- Literature comprehension: Provides key passages from research papers and evaluates whether the model can correctly restate, complete, or generalize the reasoning chain.
- Construction and computation: Requires the model to construct counterexamples, provide concrete instances, or complete mathematically verifiable calculations. The expected answer format is explicit and can be checked automatically.
- Exploratory proving: Focuses on research-level proof tasks and most closely resembles the actual form of scientific work. This category is intended to probe the upper boundary of a model’s ability to work on problems without a standard solution path.
This classification is valuable because it shows where a model is strong or weak instead of reducing its performance to a single overall score.
For example, a model might perform well on knowledge retrieval, making it a capable literature-review assistant, while performing poorly on exploratory proving. That would indicate that it remains far from independently conducting original research.
5. What the Results Mean: A Reality Check and a Roadmap
The evaluation’s main findings can be summarized in three points:
- No breakthroughs on open problems: None of the evaluated models independently solved an unresolved open question. This indicates that a fundamental gap remains in their ability to perform original reasoning.
- Reliability improves with verification: On tasks whose answers can be checked by machines, model scores are more trustworthy and reproducible. This provides a methodological template for future evaluation systems.
- Research ability is highly granular: Models perform very differently across research activities. “AI research capability” is therefore not a single metric, but a collection of abilities that must be measured separately.
For practitioners, the message is clear: instead of focusing only on leaderboard scores, examine how models perform on tasks with verifiable outcomes.
Teams considering the use of AI in research and development can apply several practical principles:
- Start with verifiable subtasks: Use AI for steps whose outputs can be checked automatically, such as assisting with formal verification, checking the validity of theorems, or judging code equivalence. Leave less verifiable creative work to humans.
- Define machine-checkable acceptance criteria: Whether the output is a mathematical proof or an engineering proposal, establish an automated validation process whenever possible. This helps prevent outputs that merely appear correct from entering production workflows.
- Use layered evaluations to define usage boundaries: Different models have different strengths. Select tools according to the task type—knowledge retrieval, comprehension, construction, or exploration—rather than relying on a single overall ranking.
6. Conclusion
Whether AI will replace mathematicians is not a realistic near-term question. A more useful question is: which parts of the research process can AI support reliably?
Machine-verifiable evaluation frameworks provide a practical way to answer that question. They replace intuition and promotional claims with checkable evidence, helping define the boundaries of current model capabilities.
For mathematicians, this may offer some reassurance. Models cannot currently solve genuinely open problems on their own. However, when used as rigorously verified assistants, they can already provide value within clearly defined limits.
As verification technology and model capabilities continue to develop, those limits will be measured and reassessed again and again.
